[ExecuTorch][WebGPU] linear_q4gsw test suite: Llama-1B shapes + 4k/8k sweep#20266
Merged
Conversation
…ation) Pull Request resolved: #20201 Backend-agnostic GPU-timestamp infrastructure, split out so the general implementation is foundational (below SDPA) while the SDPA-specific dispatch labeling stays above the SDPA op. Composed of: `WebGPUQueryPool`, a faithful re-port of Vulkan's `vkapi::QueryPool` (`backends/vulkan/runtime/vk_api/QueryPool.{h,cpp}`) — same `ShaderDuration` data model and ticks->ns conversion; three deviations are forced by the WebGPU API (per-dispatch bracketing via a compute-pass `timestampWrites` descriptor since there is no mid-encoder `writeTimestamp`; readback via `resolveQuerySet` + buffer map rather than host-side `vkGetQueryPoolResults`; the `TimestampQuery` capability requested as an explicit device feature, fail-open if the adapter lacks it). `WebGPUDevice` gains timestamp-feature detection, and `WebGPUGraph` gains a per-dispatch `kernel_name` label plus `execute()` bracketing of each compute pass when the pool is active. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical. The SDPA per-kernel labeling lives in the companion "for SDPA" diff above the SDPA op. Co-authored with Claude. ghstack-source-id: 392975889 @exported-using-ghexport Differential Revision: [D108188287](https://our.internmc.facebook.com/intern/diff/D108188287/)
… input_pos Pull Request resolved: #20086 Adds the fused `sdpa_with_kv_cache` op (QK attention-weights, softmax, attention-output sub-kernels over the KV cache), composing the three enablers below it: the base graph's inter-dispatch buffer passing (scratch buffers + multi-pass execute), the `update_cache` op, and the SymInt live-scalar mechanism. The QK/softmax/AV kernels mirror the Vulkan reference's flat-index/GQA/causal-mask math (NCHW, buffer-only, fp32). `input_pos` is consumed dynamically via the SymInt mechanism: the op reads `symint_buffer()` as a uniform, sizes its scratch + dispatches for the max context length, and registers a resize hook so a single delegate runs an autoregressive decode loop (feed only the new token + advancing `input_pos`) instead of a fixed baked position. Mirrors the Vulkan SymInt = live uniform-buffer design. Tests live in the stacked test-suite diff above (clean op diff here). Authored with assistance from Claude. ghstack-source-id: 392609088 @exported-using-ghexport Differential Revision: [D107595125](https://our.internmc.facebook.com/intern/diff/D107595125/)
…-graph KV cache Pull Request resolved: #20087 Adds the WebGPU SDPA test coverage as its own diff, stacked on the SDPA op (which already carries the dynamic-`input_pos` consumption) and the SymInt mechanism below it: multi-step prefill->mt->decode replay, runtime-dynamic `input_pos` (autoregressive decode), and an in-graph mutable KV cache, each compared against a torch `F.scaled_dot_product_attention` golden. - `test/ops/sdpa/test_sdpa.py`: `ReplaySeq`/`REPLAY_SEQS` + per-step replay export/golden; `DynamicSdpaModule` + `export_dynamic_decode` (one `.pte`, `input_pos` supplied at runtime as a SymInt); `DecodeCacheModule` + `export_incache_decode` (KV cache as `register_buffer` mutable buffers, so the cache persists in-graph and forward() feeds only the new token + `input_pos`). - `test/test_webgpu_native.cpp`: `test_sdpa_replay`, `test_sdpa_dynamic_decode` (+ negative control: a pinned `input_pos` diverges), `test_sdpa_incache_decode` (+ static control: a fresh Module per step diverges, proving in-graph accumulation is real), `test_symint_roundtrip`, `test_resize_hook`; shared per-element tolerance `sdpa_within_tol` (abs 1e-4 OR rel 1e-3). - `test/test_build_webgpu.sh`: export the replay / dynamic / in-graph-cache models for the native test. Authored with assistance from Claude. ghstack-source-id: 393014582 @exported-using-ghexport Differential Revision: [D107595144](https://our.internmc.facebook.com/intern/diff/D107595144/)
…near_q4gsw) Pull Request resolved: #20226 Adds the `et_vk.linear_q4gsw` operator (4-bit groupwise-symmetric weight-only linear) to the WebGPU backend: dequantize the packed int4 weight in WGSL (`(q-8)*scale`) and accumulate an fp32 matmul, consuming the serialized `[N, K/2]` uint8 weight directly (no prepack), one workgroup per output row. Mirrors the Vulkan reference (`backends/vulkan/.../impl/QuantizedLinear.cpp`). The dispatch carries a `linear_q4gsw` label for GPU-timestamp-query profiling (mirroring the SDPA kernels). The numerical test suite is in the stacked test diff. ghstack-source-id: 392908894 @exported-using-ghexport Differential Revision: [D108312283](https://our.internmc.facebook.com/intern/diff/D108312283/)
… sweep Pull Request resolved: #20227 Adds the numerical test suite for `et_vk.linear_q4gsw` (stacked on the op diff), mirroring the SDPA test suite. A named CONFIGS sweep covers real Llama-3.2-1B linear shapes — q/o-proj (2048->2048), k/v-proj (2048->512), gate/up-proj (2048->8192), down-proj (8192->2048), lm_head (2048->128256) — plus 4k/8k large-token prefill (M=4096/8192 on the 2048->2048 and 2048->512 projections). `test/ops/quantized_linear/test_quantized_linear.py` exports each config's `.pte` + an fp64 dequant-matmul "truth" golden; `test/test_webgpu_native.cpp` reconstructs the deterministic ramp input bit-for-bit, runs the op on the GPU, and compares per element; `scripts/test_webgpu_native_ci.sh` wires the fixtures into the Dawn(Tint)+SwiftShader CI. ghstack-source-id: 392908895 @exported-using-ghexport Differential Revision: [D108314849](https://our.internmc.facebook.com/intern/diff/D108314849/)
meta-cla
Bot
added
the
CLA Signed
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20266
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
JulianCloudNTH
approved these changes
Jun 13, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #20227 by @JulianCloudNTH
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/24/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/24/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/main
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/JulianCloudNTH/24/orig
@diff-train-skip-merge